AI, Trend or Hype
From innovation to reliable application in practice
At Technolution, we are always looking for ways to solve complex problems as simply as possible.When it comes to this, Artificial Intelligence (AI) is opening up a fascinating new world. There’s a tremendous range of possibilities: from smart traffic cameras to optimizing energy management. But not every breakthrough touted in the media proves useful in practice. The trick is to translate possibilities into reliable, predictable, and controllable applications.
So how do we turn the hype into real added value at Technolution?
Four core areas of applying AI
Success lies not in one algorithm or tool, but in an integrated approach: the smart combination of technologies, domain knowledge, and robust engineering. Let’s look more closely at four fields of application: Vision, Generative AI, Forecasting, and Control.
Vision: Smart image processing
The breakthrough of AlexNet in 2012 heralded a revolution in the field of image processing: ever since, convolutional neural networks have been performing better than classical methods for image recognition and classification. Neural networks have since become indispensable in production systems.
A good example of an integrated approach is the FlowCube, our smart traffic sensor. This recognizes vehicles, cyclists, and pedestrians locally, in the camera (“on the edge”), so that privacy-sensitive data never even leaves the camera: this is privacy by design. The models are based on the popular YOLO model, but thanks to extensive finetuning on the basis of actual data, our results are even better, even under challenging circumstances such as images at night.
We’re also developing our own models, for instance for object reidentification between multiple FlowCubes, which allows us to estimate travel times for cyclists between various locations, without requiring any personal data. We also apply this knowledge in the industry and the shipping industry, for example to automatically measure the draught of ships. The strength here again is in the combination of classical algorithms with modern machine learning.
Generative / Knowledge AI:
New creativity and data
With the advent of ChatGPT, large language models (LLMs) have become hugely popular: they generate text, summaries, and even code. But there are drawbacks: hallucinations, limited transparency, and dependency on large cloud suppliers. Our way of dealing responsibly with this is to combine the power of AI with controllability. Retrieval-Augmented Generation (RAG), for instance, allows output to be traced back to its source data, so that the information can be verified. And we use open-weight models to build secure, cost-efficient solutions that also work on-premise: independent of the cloud and with a lower risk of vendor lock-in.
We can compensate for the inherent uncertainty of language models by looking for solutions that involve human beings in the process. For example, by making automatic text proposals to describe changes in the configuration, which means less work and simpler version management.
The technology underlying language models is generative AI; essentially this learns probability distribution across a huge number of examples. The examples can be text, as in the case of large language models, but also image, video, sound, and much more. There are countless potential applications: from generating text to generating time series of energy consumption, to give a different example. Due to privacy legislation, you can’t simply use data from smart meters, but this technology makes it possible to simulate fictitious but realistic time series.
Forecasting: Reliable predictions
Forecasting time series, for example for motorized traffic or energy consumption, is experiencing somewhat less of a boom. The interesting aspect is the distinction between model-driven approaches and self-learning approaches. If you use existing knowledge and models, you retain greater control, but AI yields better results in some domains simply by looking at the data.
For instance, what we’re seeing in experiments with graph neural networks for traffic forecasting is that performance improves when the neural network learns the dynamic of the road network by itself on the basis of speed and intensity measurements, compared to when we impose the structure beforehand. The manual model is too simplistic; modern self-learning systems are capable of understanding more complex patterns.
The difference here between academic research and industrial research is significant. Scientists often optimize for abstract metrics, such as the average speed on a road section. In practice, however, requirements can be quite different, for example if we want to find out where traffic jams emerge and how they disappear again. In this case, the exact speed is not relevant. Questions like these require a different approach, and this is why we adapt the neural network models to our clients’ actual questions. This is how we ensure our predictions are not only accurate but also relevant.
Control: Optimization and control
New technologies such as deep reinforcement learning are very promising and the science is developing in leaps and bounds. In practice, however, things remain challenging. That’s why, for critical systems, we continue to rely on proven methods such as model-predictive control and genetic algorithms.
And yet we keep a close watch on reinforcement learning, because this is likely to become very significant over the coming few years. Take traffic lights: the FlowCube provides vision, MobiMaestro does the forecasting and control. The challenge here is that safety always comes first: AI must never create dangerous situations. This is why we are building a protective framework around the AI.
AI is also opening doors in energy management systems. Smart algorithms determine how a battery and solar panels can be used optimally, depending on the energy demand and the weather forecast. But this involves a degree of uncertainty, and we mitigate this by using stochastic optimization – not just point estimates but also the reliability of a predication. This is how we build solutions that are robust.
Creating value with AI
AI should not just be smart, but above all comprehensible, reliable, and applicable in practice. This doesn’t change our principle that systems should be as simple as possible. Unfortunately, AI is no miracle cure for every problem. Real added value only emerges when it is combined with domain knowledge, robust engineering, and critical reflection – and these are things we at Technolution happen to be good at.
